MarvinChung
commented
1 year ago Many modifications have been made to the code, and I'd like to highlight some specific ones.
In the train section, there are two key processes:
Training Preparation for Nerf (training_prep_nerf): This process involves preparing the density grid and handling the density_grid_sample_ct, which stores the sample count for each voxel.
Nerf Training (train_nerf): This step includes both sampling and training. When a ray traverses a voxel that is not considered empty, we perform point sampling on the ray's position and store this information for later loss computation. During the loss computation, we examine the sampled points. If the weight exceeds a certain threshold, we classify the point as part of the surface. Subsequently, we increment the voxel count for ray casting triangulation.
This is a very good work. I am currently looking at the ray casting triangulation part in the code, can you tell me which part of the original Instant-NGP code you modified, and be specific? Thanks!